How should AI-generated work be checked before publication?

Publishing changes the calculus in two ways. The output leaves the building, so an error cannot be corrected before someone acts on it. And it acquires your name, so a fabricated statistic is not an internal mistake but a claim your business made. Both argue for a check that is targeted rather than general, because a general read is exactly the check that fluent text passes.

The first target is every specific. Numbers, dates, names, citations, quotations, product details, statutory references, links. Each of these is a place where the shape of the sentence demanded a particular value, and where an unavailable value gets supplied anyway. Marking them and checking only those is faster than reading carefully and finds substantially more. A specific that cannot be traced to a source should be removed rather than softened, since a hedged fabrication is still a fabrication.

The second is every commitment. Anything the text promises on the business's behalf: a timescale, a price, an availability, a guarantee, a capability. These are not accuracy questions but authority questions — the text may be entirely plausible and still commit you to something nobody agreed. They are easy to spot once you look for them and easy to miss when reading for sense, because they read as ordinary confident copy.

The third is duplication against your own existing material. Generated text on a familiar subject tends toward the same structure and the same points, so a business publishing regularly accumulates pages that differ mainly in wording. This is a real problem for search visibility and a larger one for readers, and it is invisible when each piece is reviewed on its own. Checking against what already exists is the only stage where it can be caught.

There is a fourth check that applies to anything making a claim about a third party: whether the claim is supportable and whether it is fair. A statement about a competitor, a supplier, a named individual or a category of business carries consequences that an internal document does not, and generated text produces these confidently because they are grammatically ordinary. This is where the cost of publishing without checking is highest and least recoverable.

Finally, the check should leave a record. Who read it, when, and what they changed. This is partly for accountability and mostly because the record is what tells you, later, whether the process is being followed when things are busy. A check that happens sometimes and is never recorded is indistinguishable from one that does not happen.